About
I am currently a researcher at the Language Model Lab, Huawei Hong Kong Research Center. I obtained my Ph.D. degree from The Chinese University of Hong Kong supervised by Prof. Michael R. Lyu and Prof. Irwin King, and the B.Eng. Degree from Yingcai Honors College of University of Electronic Science and Technology.
Our team's effort is on large language models with topics spanning from pre-training, post-training, to agentic AI (e.g., deep research and coding agent). I am also an experienced researcher in LLM efficiency, e.g., compression and acceleration of LLMs.
News
-
2026-8
🔥 We are happy to announce Lego-RL, an open-source framework for harness-native reinforcement learning of coding agents. All code, data, and models are available on
GitHub and Hugging Face.
-
2026-8
🔥 We establish
LegoX, an open-source collection spanning agentic data pipelines, training frameworks, evaluation benchmarks, and foundation models to advance frontier agentic intelligence. LegoX has accumulated over 200K downloads on Hugging Face. See our work on GitHub and Hugging Face.
- 2026-1 We present SWE-Lego , the state-of-the-art supervised fine-tuning method for software issue resolving. All code, data, models are now opensourced. Project website.
- 2025-11 We will present the tutorial "Efficient Inference for Large Language Models – Algorithm, Model, and System" at EMNLP 2025. Tutorial website.
Selected Publications
*: Equal contribution; #: Corresponding author; +: Project lead
Invited Talks
- "Quantization and Pruning of Large Language Models: Challenges, Techniques and Opportunities" at SLAI, 2025. [Slide]
- "Efficient Inference for Large Language Models – Algorithm, Model, and System" at EMNLP Tutorial, 2025. [Tutorial website]
- "Quantization and Pruning of Large Language Models: Challenges, Techniques and Opportunities" at LMG, 2025.
Projects
LegoX: Open-Source Collection for Frontier Agentic Intelligence
LegoX is an open-source collection spanning agentic data pipelines, training frameworks, evaluation benchmarks, and foundation models to advance frontier agentic intelligence.
LegoX Projects
Lego-RL trains coding agents with online reinforcement learning inside native harnesses and real repository environments.
SWE-Review closes the loop on issue resolution with agentic code review, helping agents inspect, critique, and improve generated patches.
Terminal-Lego studies what makes interaction trajectories effective for training terminal agents across realistic command-line tasks.
SWE-Lego is a supervised fine-tuning framework for software issue resolving, built with open data, models, and reproducible training recipes.
PocketFlow automatically searches for optimal model compression strategies such as network pruning, quantization, knowledge distillation with little human efforts, and also supports TFLite deployment on Android devices. It has collected 2600+ stars and 480+ forks.